Building a Personal Knowledge Base

Appearance

Power Tips and Advanced Patterns ​

~8 minutes. Custom skills, community marketplace, CLAUDE.md mastery, and fixing common issues.

Once you're comfortable with the daily workflow, these tips will help you get more out of the system.

Creating Custom Skills ​

After Claude does something useful — formats a brief a specific way, runs a workflow you like — you can turn it into a reusable skill.

"That was great. Turn what we just did into a reusable skill."

Claude will create a markdown file in .claude/skills/ with the right structure.

Skill Anatomy ​

Every skill is a markdown file with three parts:

.claude/skills/weekly-risks.md
---
# 1. FRONTMATTER — when to trigger
name: weekly-risks
description: Summarize risks from this week's meetings
trigger: weekly risk summary, what risks came up
---

# 2. INSTRUCTIONS — what to do
Read all meeting notes from this week in the Meetings/ folder.
Extract any mentioned risks, concerns, or blockers.
Group them by project.

# 3. OUTPUT FORMAT — how to present it
Format as:
## Weekly Risk Summary — [date range]
### [Project Name]
- **Risk**: [description]
  **Raised by**: [[person]] | **Severity**: high/med/low
  **Mitigation**: [if discussed]

That's it. No special syntax, no compilation, no registration. Save the file and the skill is live.

When to Create a Skill ​

The best skills come from workflows you've already done manually 3+ times. If you find yourself typing the same kind of prompt repeatedly, it's time to make a skill.

Good candidates:

  • Weekly risk summaries
  • Sprint retrospective synthesis
  • Stakeholder briefing for specific audiences
  • Onboarding summaries for new team members
  • Project handoff documents

Installing Community Skills ​

Skills can be shared within teams or communities. To install a skill someone else wrote:

Browse Available Skills ​

Ask Claude what's available:

"What skills are available to install? Show me what's in the skill catalog."

Or browse your team's shared skill repository if one exists. Skills are just markdown files — they can live in any shared folder, Git repo, or wiki.

Install via CLI or Conversation ​

bash
# If your setup has a plugin/skill marketplace
claude plugin install <skill-name>

# Or just ask Claude
# "Install the meeting-prep skill from the team catalog"

You can also install manually: copy the .md file into your .claude/skills/ directory.

After Installing ​

  • Read the skill file to understand what it does
  • Test it with a sample input
  • Modify it if the output format doesn't match your preferences

Contributing Skills ​

If you've built a useful skill, share it with your team:

  1. Polish the skill — clean up instructions, add good trigger phrases, document the output format
  2. Test with a colleague — have someone else try it without your context to see if it's clear
  3. Submit to your team's catalog — the exact process depends on your team's setup (shared repo, wiki, skill registry)
  4. Include a README — what it does, when to use it, what KB structure it expects

The best contributed skills are specific enough to be useful but general enough to work across different people's KB structures.

CLAUDE.md Best Practices ​

Your CLAUDE.md is the most important file in your workspace. It's the persistent instruction set that shapes every interaction.

DO Include
✅ Your role and team — "I'm a PM on the Platform team"
✅ Output preferences — "Use bullet points, not paragraphs"
✅ KB folder paths — "Meeting notes go in My_KB/Meetings/"
✅ Extraction rules — "Always extract action items with owners"
✅ Corrections — "George = George Costanza, not George Washington"
✅ Naming conventions — "Files: YYYY-MM-DD-topic-slug.md"
DO NOT Include
❌ Skill instructions — those belong in skill files
❌ API keys or secrets — use environment variables
❌ Huge data dumps — eats up context window
❌ Contradictory rules — Claude picks one arbitrarily
❌ Things that change daily — use the Work Board instead
❌ Entire org charts — let people files handle this

Maintenance Rhythm ​

Review your CLAUDE.md monthly. Ask Claude:

"Read my CLAUDE.md and tell me if anything looks outdated or contradictory."

Claude will flag issues. Update as your role, projects, and preferences evolve.

Context Window Management ​

Claude's context window is large but finite. Here's how to manage it effectively.

Signs the Context Window Is Full ​

  • Claude starts giving shorter, less detailed answers
  • It "forgets" something you mentioned earlier in the conversation
  • You see an explicit message about context limits
  • Responses become more generic and less personalized

Strategies ​

1
Use /compact to compress context
Claude summarizes the conversation so far, freeing up space while retaining key information. Use this when you're mid-task and don't want to start over.
2
Use /clear for a fresh start
Clears the entire conversation. Use this between unrelated tasks. Your KB files and CLAUDE.md persist — you don't lose your setup.
3
Break work into sessions
Process one meeting per session. Write one brief per session. Don't try to do everything in a single marathon conversation.
4
Reference files by path instead of pasting content
"Read People/Elaine-Benes.md" is better than pasting the entire file into the chat. Claude reads the file on demand and only loads what it needs.

Debugging Common Issues ​

Skill Not Matching ​

Symptom: You say "process this meeting" and Claude handles it conversationally instead of using the /process-meeting skill.

Fix:

  • Use the slash command directly: /process-meeting
  • Check that the skill file exists in .claude/skills/
  • Verify the trigger phrases in the skill's frontmatter match what you're saying
  • Restart Claude Code — skills are loaded at startup

MCP Connection Failed ​

Symptom: Claude says it can't reach a tool, or a search returns nothing.

Fix:

  • Ask Claude: "Check if my search tool is connected"
  • Look for error messages in the Claude Code output panel
  • Restart Claude Code to re-establish MCP connections
  • Check your network connection (VPN, proxy, firewall)
  • Verify credentials haven't expired

Generic Answers (Not Using KB Context) ​

Symptom: Claude gives a generic response instead of pulling from your KB files.

Fix:

  • Make sure you're in the right workspace (the one with your My_KB/ folder)
  • Point Claude to specific files: "Read People/George.md and then answer my question"
  • Check that your CLAUDE.md references the correct KB paths
  • If the conversation is long, use /compact — Claude may have lost context

Extraction Missed Something ​

Symptom: A processed meeting is missing an action item, decision, or position you know was discussed.

Fix:

  • This is the 80% rule in action. Review and add what's missing manually.
  • Tell Claude: "You missed George's action item about the platform sign-off. Add it."
  • If it's a recurring miss, update your CLAUDE.md with an extraction rule: "Always extract action items even when they're implied, not explicitly stated."
  • Consider that the transcript quality matters — mumbled words, cross-talk, and sidebar conversations may not capture well.

Files in the Wrong Location ​

Symptom: Claude saved a file to the wrong folder, or used the wrong naming convention.

Fix:

  • Tell Claude: "That file should be in Meetings/, not 0-Inbox/. Move it and update CLAUDE.md so you remember."
  • Claude will move the file and update its own instructions. This is the "living document" approach — corrections make the system better over time.
That's the guide.
Start with the daily rhythm. Process your meetings. Let the knowledge compound.
The first week is useful. After a month, it's indispensable.

Back to: Home

Built with VitePress